Measurement of Keyholes and Capacities in Multiple-Input Multiple-Output (MIMO) Channels

MITSUBISHI ELECTRIC RESEARCH LABORATORIES http://www.merl.com Measurement of Keyholes and Capacities in Multiple-Input Multiple-Output (MIMO) Channels Almers, P.; Tufvesson, F. TR23-4 August 23 Abstract It has been predicted theoretically that for some environments, the cacity of wireless MIMO systems can become very low even for uncorrelated signals; this effect has been termed eyhole or pinhole. In this paper the first measurements of this effect are presented. The measurements were performed in a controlled indoor environment that was designed to create a eyhole effect. We analyze limitations for measurement-based capacity calculations and eyhole investigations. We further present error bounds for the capacity and eigenvalue distributions due to measurement imperfections such as finite signal-to-noise ratio and multipath leaage. The bounds are compared to the measurement results and show excellent agreement. This wor may not be copied or reproduced in whole or in part for any commercial purpose. Permission to copy in whole or in part without payment of fee is granted for nonprofit educational and research purposes provided that all such whole or partial copies include the following: a notice that such copying is by permission of Mitsubishi Electric Research Laboratories, Inc.; an acnowledgment of the authors and individual contributions to the wor; and all applicable portions of the copyright notice. Copying, reproduction, or republishing for any other purpose shall require a license with payment of fee to Mitsubishi Electric Research Laboratories, Inc. All rights reserved. Copyright c Mitsubishi Electric Research Laboratories, Inc., 23 2 Broadway, Cambridge, Massachusetts 239

MERLCoverPageSide2

Publication History:. First printing, TR-23-4, August 23

Measurement of Keyholes and Capacities in Multiple-Input Multiple-Output (MIMO) Channels Peter Almers,2, Student Member, IEEE, Fredri Tufvesson and Andreas F. Molisch,3 Senior Member, IEEE Dept. of Electroscience, Lund University, Box 8, SE-22 Lund, Sweden. 2 Telia Research AB, Box 94, SE-2 2 Malmö, Sweden. 3 Mitsubishi Electric Research Labs, 8 Central Avenue, Murray Hill, NJ 7974, USA. Email: {Peter.Almers, Fredri.Tufvesson, Andreas.Molisch}@es.lth.se Abstract It has been predicted theoretically that for some environments, the capacity of wireless MIMO systems can become very low even for uncorrelated signals; this effect has been termed eyhole or pinhole. In this paper the first measurements of this effect are presented. The measurements were performed in a controlled indoor environment that was designed to create a eyhole effect. We analyze limitations for measurement-based capacity calculations and eyhole investigations. We further present error bounds for the capacity and eigenvalue distributions due to measurement imperfections such as finite signal-to-noise ratio and multipath leaage. The bounds are compared to the measurement results and show excellent agreement. Index Terms Keyhole, MIMO, measurements, eigenvalue, double complex Gaussian, pinhole, capacity. I. INTRODUCTION Multiple-input multiple-output (MIMO) wireless communication systems are systems that have multi-element antenna arrays at both the transmitter and the receiver side. It has been shown that they have the potential for large informationtheoretic capacities, since the system can provide several independent communication channels between transmitter and receiver []. In an ideal multipath channel, the MIMO capacity is approximately N times the capacity of a single-antenna system, where N is the smaller of the number of transmit or receive antenna elements. Correlation of the signals at the antenna elements leads to a decrease in the capacity - this effect has been investigated both theoretically [2][3] and experimentally [4]. It has recently been predicted theoretically that for some propagation scenarios, the MIMO channel capacity can be low (i.e., comparable to the SISO capacity) even though the signals at the antenna elements are uncorrelated [][6]. This effect has been termed eyhole or pinhole. It is related to scenarios where rich scattering around the transmitter and receiver lead to low correlation of the signals, while other propagation effects, lie diffraction or waveguiding, lead to a ran reduction of the transfer function matrix. Several previous measurement campaigns have searched for the eyhole effect due to corridors, tunnels, or diffraction in real environments, but the effect has been elusive and, to our nowledge, no measurements of a eyhole have been presented in the literature. We are using in this paper the original definition of eyholes. Recently, some authors have called eyhole any scenario that shows a reduction of the ran of the transfer function matrix (compared to the i.i.d. complex Gaussian case). This definition would imply that any scenario with strong correlation (small angular spread) is a eyhole. In this paper, we present the results of a measurement campaign that for the first time shows the eyhole effect experimentally. The measurements were performed in a controlled indoor environment, where the propagation from one room to the next could only occur through a waveguide or a hole in the wall. The measurement results show almost ideal eyhole properties; the capacity is low, the ran of the transfer matrix is nearly one though the correlation between the antenna elements is low. The experiments were based on a theoretical analysis of the requirements on both the measurement parameters and the considered environment, for finding eyholes in MIMO measurement. We present this analysis, error bounds for the capacity and eigenvalue distributions due to measurement imperfections. II. CAPACITIES AND KEYHOLES With the assumption of flat fading, we use the conventional MIMO model for the received signal vector as y = Hx + N, () where h Hi is the channel transfer matrix normalized as, E H 2 F = N R N T, x is the transmitted signal vector, and N CN, σ 2 n I represents noise. For equal power allocation between the transmitter elements the Shannon channel capacity [bit/s/hz] for the channel model in () can be calculated as [] C = µ log 2 µdet I + γ eval HH N T (2) = KX µ log 2 + γ eval (H) s, N T (3) = where [ ] denotes the complex conjugate transpose, s (H),is the :th singular value of H and γ eval is the signal-to-noise ratio (SNR) that the capacity is evaluated for. The latter is defined as γ eval = E h mm h mm h i = σ2 h E n mm n σ 2 = mm n σ 2, (4) n where E [ ] is the expectation, h mm and n mm are the mm entry in H and N respectively, σ 2 h is the variance of the entries in H. It is assumed that x has unity energy.

2 It is well nown that correlation between the antenna elements at the transmitter and receiver reduces the capacity. The channel transfer matrix is often modeled as [2][7][8]: H = R /2 R GRT/2 T, () where R /2 R and R /2 T describes the correlation between the signals at receiver and transmitter elements, respectively, and G is a matrix with i.i.d. complex Gaussian entries, G CN (, I). The square root is defined as R /2 R /2 = R. According to this model, the channel, and thus the capacity, is completely determined by the correlations at transmitter and receiver. However, in [] it is shown that a more general model is H = R /2 R G RT /2 G T R T/2 T, (6) where G R and G T are both i.i.d. complex Gaussian matrices, and where T /2 describes the transfer matrix between the transmitter and receiver environments. For a perfect eyhole, e.g. a single-mode waveguide [6] between the transmitter and the receiver environment, T /2 has ran one. This results in a total channel transfer matrix, H, of ran one as well, even though R /2, G R and G T have full ran. R, R/2 T III. MEASUREMENT SETUP The measurements were performed with one antenna array located in a shielded chamber, and the other array in the adjacent room. A hole in the chamber wall was the only propagation path between the rooms. We measured three different hole configurations: ) A hole of size 47 22 mm with a 2 mm long waveguide attached (referred to as waveguide ). 2) A hole of size 47 22 mm without waveguide ( small hole ). 3) A hole of size 3 3 mm without waveguide ( large hole ). The large hole is intended as test measurement, in which we do not anticipate a eyhole effect. Linear virtual arrays with Fig.. [mm] 4 42 6 27 Networ analyzer 3 3 Overview of the measurement setup. 6 Absorbing material Shielded chamber 67 3 6 antenna positions and omnidirectional conical antennas are used both at the transmitter and the receiver; the measurements were done during night time to ensure a static environment. In 4 Fig. an overview of the measurement setup is presented. The measurements were performed using a vector analyzer (Rohde &SchwarzZVC)at3. 4. GHz, where M =complex transfer function samples spaced MHz apart were recorded. The received signal is amplified by 3 db with an external low noise power amplifier to achieve a high signal-to-noise ratio (SNR). The measurement SNR was estimated to 26 db, where in this case noise include thermal noise, interference and channel changes during the measurements. The measurements yielded the elements of the channel matrix H, and via Eq. 2, the outage capacity. We obtained samples of H, and a capacity realization corresponds to each of those realizations; this allow plotting a cumulative distribution of the capacity and the determination of the outage capacity []. In order to compare the capacity reduction due to correlation to the capacity reduction due to the eyhole effect, we first estimate the correlation matrices ˆR T and ˆR R from the measurements as MX ˆR T = H T MN m H m, (7) T m= MX ˆR R = H m H MN m, (8) R m= and then we use the Kronecer model in Eq. (). As an example of the spatial correlation of the waveguide measurements we here present the first column of ˆR T and ˆR R ˆr T =..4.87.22.46.9 T (9) ˆr R =..42.37.24.7.8 T () These values were obtained with a limited number of transfer function samples (M =) due to the long duration of the measurements. The correlation matrices allow a prediction of the capacity decrease due to signal correlation (i.e., an effect that is different from a eyhole effect). Inserting those results into Eq. () gives the outage capacity in a correlated, noneyhole channel, where the correlation matrix corresponds to measured correlations. IV. CAPACITY ANALYSIS AND RESULTS In this section we analyze the influence of using nonideal transfer function measurements when calculating the capacity, and analyze their effect on our experiment. When measuring a transfer matrix, the measured quantity will consist of contributions not only from multi-path components (MPC) but also from measurement noise. For eyhole measurements, the measured transfer matrix will in addition to the noise and the eyhole transfer matrix also consist of MPC leaage, as described below. The measured eyhole transfer matrix can be modelled as H meas = H ey +H lea + Ñ, () where Ñ CN, σ 2 ñ I denotes the noise. For an ideal eyhole the eyhole transfer matrix has ran one with entries belonging to a double complex Gaussian distribution []. The eyhole transfer matrix can therefore be modeled as H ey = g R g T, where g R, g T are column vectors

% outage capacity [bit/s/hz] % outage capacity [bit/s/hz] 3 g R, g T CN, σ hey I. The MPC leaage describes MPCs propagating between the transmitter and the receiver via other paths than through the eyhole. The leaage matrix is modeled as Gaussian and as a worst case assumption, in terms of measuring a eyhole, it is assumed to have independent entries, H lea CN, σ 2 lea I. We thus see that noise leads to the same destruction of the eyhole effect as leaage. In order to chec the effect on the evaluated capacity, we lump the two effects into a single matrix, Ň = H lea + Ñ. Assume that H ey, H lea and Ñ are independent, the measurement SNR (including leaage) from a eyhole measurement point of view, γ meas, can therefore be defined as γ meas = σ2 h ey σ 2 (2) ň h i E h meas (m, m)h meas(m, m) = i. E hh lea (m, m)h lea (m, m) + E [ñ(m, m)ñ (m, m)] In Fig. 2 simulations of the % outage capacity for three different measurement SNRs, versus the evaluated SNR are plotted [9]. Additionally we plot the capacity for our eyhole measurement, which, as mentioned before, has an estimated measurement SNR of 26 db. From the figure it In Fig. 3 the measured % outage capacities versus the number of antenna elements, N R = N T, are shown for an SNR of db. For comparison, the figures also presents the i.i.d. capacity, the correlated capacity and the capacity for a perfect theoretical eyhole. We see that the measured % outage capacity for the waveguide setup is very close to the simulated perfect eyhole. With the large hole the capacity isclosetoani.i.d.channelanditsmeasuredcapacityisin between the curves for the i.i.d. model and the correlated model. The difference in % outage capacity between the measured waveguide and large hole configuration is up to 7 bit/s/hz. The correlation model shows the decrease in capacity related to the receive and transmit antenna correlation, and shows a capacity that is more than bit/s/hz higher than the waveguide measurement. In this figure it can clearly be 2 2 i.i.d. large hole correlation model small hole waveguide perfect eyhole 2 2 = 2 db = 2 db waveguide γ meas = 26 db = 3 db = db 2 3 4 6 Array size Fig. 3. % outage channel capacity vs. the antenna array size, where N R = N T. 2 2 3 Evaluation SNR [db] Fig. 2. The estimated capacities for different simulated measurement SNRs {2, 2, 3, } versus evaluation SNR [db] togehter with the measured eyhole with a estimated eyhole SNR of 26 db. can be concluded that the SNR of the eyhole measurement has to be around db better than the SNR used in the evaluation to give capacity values close to the ideal case. ForanevaluationSNRof db the contribution from noise and leaage components has to be 2 db lower than the contributions from the eyhole. This can be difficult to find for real-life eyhole situations, e.g. tunnel waveguiding or diffraction, which might explain the difficulty in finding eyholes in previous measurement campaigns. The shielded chamber in our measurements made sure that the leaage was very small. The SNR (including leaage) was estimated to 26 db. Therefore an evaluation SNR of db is appropriate in this case. seen that the capacity for the waveguide setup nearly follows that of a perfect eyhole. The increase in capacity for more antenna elements is due to the antenna gain of the receiver array. The capacity of the large hole increases almost as the capacity for the i.i.d. channel. This shows that the eyhole effect has disappeared entirely when the large hole of size 3 3 mm is the (only possible) path between transmitter and receiver. V. SINGULAR VALUE ANALYSIS AND RESULTS The ran of an ideal eyhole is one, and in order to characterize a measured eyhole it is therefore of interest to study the distribution of the singular values (or eigenvalues). Especially, the ratio between the largest and second largest singular values should be large. For a measured eyhole, however, this difference is dependent on the measurement SNR of the eyhole, γ meas. Again we consider the leaage as part of the noise, and add the leaage and the measured noise together as, Ň = H lea + Ñ. The difference between the :th singular value of the measured transfer matrix (), denoted s (Hey+Ň), and the :th singular value of the eyhole transfer matrix,

Mean eigenvalue [db] Mean eigenvalue [db] 4 s (Hey), is limited by [] s(hey+ň) s (Hey) s(ň), (3) where the singular values for the ran one eyhole matrix, H ey,are ( s (Hey) >,= s (Hey) =,>. (4) From (3) and (4) the second largest singular value of the measured eyhole transfer matrix can now be upper bounded as s (Hey+Ň) 2 s (Ň). () With the variable transformation, γmeas ˆN = Ň, where ˆN is normalized as ˆN 2 = N RN T, the second largest singular F value of the measured eyhole is upper bounded as a function of the measurement SNR. The largest singular value of the normalized noise matrix s (H meas) 2 s ( ˆN). (6) γmeas This means that if the second largest singular value of the measured transfer matrix exceeds the threshold (6), we do not have a eyhole with ideal properties. The expectation, E γmeas s ( ˆN), of the largest singular value of ˆN can be found from the density function of ordered eigenvalues [] of the semi definite Wishart matrix ˆN ˆN, since those eigenvalues are equal to the magnitude squared of the singular values of, ˆN, therefore where E s (H+Ň) 2 s E λ ( ˆN ˆN) = E γmeas s γmeas s ( ˆN) E λ ( ˆN ˆN) (7), (8) s Z = λ f Λ (λ ) dλ, (9) and the marginal density function of the largest eigenvalue, f Λ (λ ), is given in []. In Fig. 4 the bound for the second largest singular value is plotted for different measurement SNRs together with simulated values of the largest and second largest singular values of an ideal eyhole with measured additive white Gaussian noise, and our measured mean values of the largest and second largest singular values. In Fig. the mean of the measured ordered eigenvalues for the different channel setups are presented. It can be clearly seen that the waveguide channel is of low ran. The difference between the mean of the largest and second largest eigenvalue is almost 3 db for the measured eyhole. As a comparison, the difference between these eigenvalues for the ideal i.i.d. channel is around 2 db. The difference between the largest and smallest eigenvalues (i.e., the condition number of the matrix H m H m) is almost db for the waveguide and around 2 db for the i.i.d. channel. 2 - - bound sim. eig. - sim. eig. 2 meas. eig. meas. eig. 2-2 - 2 2 3 Measurement SNR [db] Fig. 4. The mean of the largest and the second largest singular value together with the upper bound for the second largest eigenvalue, both as a function of the measured SNR. The waveguide measurement with an estimated measurement SNR of γ meas 26 db is also plotted. Fig.. 2 - -2 i.i.d. -3 large hole correlation model small hole waveguide -4 2 3 4 6 Ordered eigenmodes Mean of the ordered singular values [db]. A. Keyhole capacity error For a eyhole measurement the capacity from a theoretical eyhole could be upper bounded in respect to measurement SNR, γ meas, using the results in (6) and Jensen s inequality as µ E[C ey ] < log 2 + γ h i 2 eval E s (Hey) (2) N T à +(K ) log 2 + γ eval E s ( ˆN) 2!. N T γ meas The discrepancy in capacity compared to a theoretical eyhole could then be upper bounded from the results in (6) as à ε (E[C ey ]) < (K ) log 2 + γ eval E N T γ meas s ( ˆN) 2!. (2) In Fig. 6 simulated errors and the error bounds are presented for different measurement SNRs and evaluation SNRs for the

% outage capacity error [bit/s/hz] PDF and histogram PDF and histogram case with K =6singular values. The bound is not that tight in the region where γ meas /γ eval < db. Fig. 6. 9 8 7 6 4 3 2 = 2 db = 2 db = 3 db bound γ meas = 2 db bound γ meas = 2 db bound γ meas = 3 db 2 Evaluated SNR [db] Capacity discrepancy from theoretical capacity of a eyhole. VI. ENVELOPE DISTRIBUTION If no eyhole is present, then the amplitudes of the entries in H follow a Rayleigh distribution. However, if a eyhole exists, the transfer matrix T /2 in (6) is the all one matrix [], and the amplitude statistics, a = h ey (m, m), followa double-rayleigh distribution. It can be shown that the PDF of the envelope distribution can be expressed as ³ a ak σ hey f A (a) = σ 2, (22) h ey where K ( ) is the modified Bessel function and σ 2 h ey = h E h (m, m) h (m, m) i. In Fig. (7) we present histograms of the received amplitudes in the experiment, both for the waveguide case and the large-hole case. As a reference we have also shown the PDFs for the amplitude of a Rayleigh variable and a double Rayleigh variable. The received amplitudes with the waveguide correspond well to the double-rayleigh distribution, which agrees with theory. The received amplitudes for the large hole, however, correspond to a Rayleigh distribution since in this case the channel can be described as one rich scattering channel though all paths into the chamber is through the large hole. VII. CONCLUSIONS In this paper the first experimental evidence of the eyhole effect in wireless MIMO systems is presented. Using a controlled indoor environment, we found a eyhole with almost ideal properties: the correlations at both the receiver and at the transmitter are low but still the capacity is very low and almost identical to a theoretical perfect eyhole. In our measurements we use a waveguide, a small hole, and a hole of size 3 3 mm as the only path between the two rich scattering environments. For the waveguide case, we found almost ideal eyhole properties, but for the large hole the capacity is almost.8 waveguide double complex Gaussian complex Gaussian.6.4.2.. 2 2. 3.8 large hole double complex Gaussian complex Gaussian.6.4.2.. 2 2. 3 Envelope of channel coefficients Fig. 7. Envelope distribution for the waveguide and for the large hole measurements. as large as for a theoretical Gaussian channel with independent fading between the antenna elements. We then presented an analysis of the sensitivity with respect to noise and alternative propagation paths. We found that both the noise and the alternative propagation paths during the channel sounding must be approximately db weaer than the noise level considered for the capacity computations. From this we can conclude that the eyhole effect due to realworld waveguides lie tunnels or corridors will usually be very difficult to measure. Acnowledgement: Part of this wor was financed by an INGVAR grant from the Swedish Foundation for Strategic Research. REFERENCES [] G. J. Foschini and M. J. Gans, On limits of wireless communications in fading environments when using multiple antennas, Wireless Personal Communications, vol.6,pp.3 33,998. [2] D. Shiu, G. J. Foschini, M. J. Gans, and J. M. Kahn, Fading correlation and its effect on the capacity of multi-element antenna systems, IEEE Transactions on Communications, vol. 48, pp. 2 3, March 2. [3] C. N. Chuah, D. N. C. Tse, J. M. Kahn, and R. A. Valenzuela, Capacity scaling in mimo wireless systems under correlated fading, IEEE Transactions on Information Theory, vol. 48, pp. 637 6, 22. [4] A. F. Molisch, M. Steinbauer, M. Toeltsch, E. Bone, and R. S. Thoma, Capacity of MIMO systems based on measured wireless channels, IEEE Journal on Selected Areas in Communications, vol. 2, pp. 6 69, April 22. [] D. Gesbert, H. Bölcsei, D. A. Gore, and A. J. Paulraj, Mimo wireless channels: Capacity and performance prediction, in Proc. GLOBECOM, vol. 2, pp. 83 88, IEEE, 2. [6] D.Chizhi,G.J.Foschini,M.J.Gans,andR.A.Valenzuela, Keyholes, correlations, and capacities of multielement transmit and receive antennas, IEEE Transactions on Communications, vol.,pp.36 368, April 22. [7] H. Bölcsei and A. J. Paulraj, Performance analysis of space-time codes in correlated rayleigh fading environments, in Proc. Asilomar Conference on Signals, Systems and Computers, pp. 687 693, IEEE, November 2. [8] D. A. Gore and A. J. Paulraj, Mimo antenna subset selection with space-time coding, IEEE Transactions on Signal Processing, vol., pp. 28 288, October 22. [9] P. Kyritsi, R. A. Valenzuela, and D. C. Cox, Channel and capacity estimation errors, in IEEE Communications Letters, pp. 7 9, December 22. [] R. A. Horn and C. R. Johnson, Topix in Matrix Analysis. London: Cambridge University Press, 99. [] I. E. Telatar, Capacity of multi-antenna gaussian channels, European Transactions on Telecommunications, vol., November December 999.